James Washington
writing about archive utils free game ideas replays
Technique

Scope heavy MCP servers to projects, not the user

Only load big AI plug-ins in the projects that use them, saving memory in every other session.

Upgrading the AI's workbench

how it works

Moved REA and screenpipe MCPs from user to local scope (~250 MB per session for a tool used once); re-register via `claude mcp add-json -s local` with jq from ~/.claude.json.

Technique: A pattern or hard-won lesson, written down so it isn't learned twice.

The code for this one lives in a private repo. This page explains the idea so you can build your own.

claude-codemcpperformancetechnique

related

  • screenpipe / screenpipe-air MCP launcher (James wrapper) MCP server Lets the AI search what was on screen or said, on either computer.
  • claude-usage: spend rolled up by project and org Command-line tool Shows how much AI usage each project and company cost.
  • rea / rea-air (remote REA over ssh stdio) MCP server Runs the reverse-engineering toolkit on the laptop so risky binaries never run on the server.
  • One-shot scheduled agent follow-up Automation A one-time scheduled AI run that checked on phone calls it had placed and reported the outcome, then deleted itself.
  • Offload heavy/risky reverse engineering to a second machine Technique Run untrusted programs and heavy analysis on the laptop, never on the server holding the passwords.
← everything I've built
© 2026 James Washington
github twitter rss